Update README.md

README.md

@@ -1,3 +1,4 @@
+---
 base_model: google/gemma-3-1b-it
 tags:
 - text-generation
@@ -8,7 +9,6 @@ tags:
 - wall-e
 - lightweight
 - mobile-friendly
-- privacy-preserving
 - local-ai
 - multilingual
 - coding-assistant
@@ -17,126 +17,165 @@ license: apache-2.0
 language:
 - en
 - fa
-- De
+- de
 library_name: transformers
 pipeline_tag: text-generation
+---
 
 [![Open in HF Space](https://img.shields.io/badge/🤗%20Try%20Live%20Demo-FFD21E?style=for-the-badge&logo=huggingface)](https://huggingface.co/spaces/sinamsv0/WALL-E-DEMO)
 [![GitHub](https://img.shields.io/badge/⭐%20GitHub-181717?style=for-the-badge&logo=github)](https://github.com/unknownmsv/WALL-E)
 [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg?style=for-the-badge)](LICENSE)
 
-# 🤖 WALL•E — Your Personal 1B Parameter AI Assistant
+# 🤖 WALL•E — Lightweight Local AI Assistant (1B)
+
+**WALL•E** is a fine-tuned, lightweight language model based on **Gemma 3 1B**, designed for **local, privacy-preserving AI usage**.  
+It focuses on *practical tasks*, *fast responses*, and *real-world utility* rather than model size.
+
+---
+
+## 🎯 Why WALL•E?
+
+Most modern AI models are either:
+- Too large to run locally, or
+- Too generic for everyday tasks
+
+**WALL•E** is built to fill that gap.
+
+✅ Runs entirely locally  
+✅ No API keys or cloud services  
+✅ Designed for low-resource environments  
+✅ Open-source and transparent  
+
+---
+
+## ✨ Key Capabilities
 
-**WALL•E** is a fine-tuned, lightweight AI model based on **Gemma 3 1B**, optimized for practical everyday tasks. Designed to run entirely locally on your device, it prioritizes **privacy, speed, and utility** over sheer size.
+### 🌐 Multilingual Support
+- **English** – primary interaction language
+- **فارسی (Persian)** – natural and fluent responses
+- **Deutsch (German)** – conversational support
 
-## 🎯 **Why WALL•E Exists**
-Most AI models are either too large for local use or too general for specific tasks. WALL•E bridges this gap by offering:
-- ✅ **Complete local execution** – No internet or API keys required
-- ✅ **Mobile & edge optimized** – Runs on Android, Linux, and resource-constrained devices
-- ✅ **Privacy by design** – Your data never leaves your device
-- ✅ **Open & transparent** – Fully open-source under Apache 2.0
+### 🛠 Practical Task Focus
+- 📝 Text summarization (articles, notes, reports)
+- 💻 Coding help (Python, JavaScript, Bash, shell)
+- 🖥 Linux command explanations & troubleshooting
+- 📚 Short factual answers and guidance
 
-## ✨ **Key Features**
-### 🌐 **Multilingual Proficiency**
-- **English**: Native-level comprehension and generation
-- **فارسی**: Fluent Persian with natural responses
-- **Deutsch**: Conversational German support
+The model is optimized to handle **short and minimal prompts** naturally (e.g. *"Hi"*, *"Explain ls -la"*), avoiding over-generation.
 
-### ⚡ **Practical Task Optimization**
-- **📝 Smart Summarization**: Condense articles, emails, documents instantly
-- **💻 Coding Companion**: Python, JavaScript, Bash, and shell scripting assistance
-- **🖥️ System Assistant**: Linux command explanations and troubleshooting
-- **🔍 Quick Information**: Factual responses without unnecessary fluff
+---
+
+## ⚙️ Technical Overview
 
-### 🚀 **Technical Excellence**
-- **Base Model**: `google/gemma-3-1b-it` with Unsloth optimizations
-- **Training**: Fine-tuned using custom datasets for safety and multilingual performance
-- **Quantization**: Available BF16 (2GB) variants
-- **Context**: 2048 token window for coherent conversations
+| Component        | Details |
+|------------------|--------|
+| Base Model       | Google Gemma 3 1B |
+| Fine-tuning      | Supervised Fine-Tuning (SFT) |
+| Framework        | Unsloth |
+| Context Length   | 2048 tokens |
+| Precision        | BF16 |
+| License          | Apache 2.0 |
 
-## 📊 **Performance Highlights**
-- **26+ downloads in first 48 hours** – Community validated
-- **<2 second inference** on mid-range smartphones
-- **40% smaller** than comparable models with similar capabilities
-- **Zero cloud dependency** – Works completely offline
+---
 
-## 🛠️ **Quick Start**
+## 🚀 Quick Start (Transformers)
 
-### **Option 1: Hugging Face Transformers**
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
 
 model_id = "sinamsv0/WALL-E"
-model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
+
 tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(
+    model_id,
+    device_map="auto"
+)
 
-pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
+pipe = pipeline(
+    "text-generation",
+    model=model,
+    tokenizer=tokenizer
+)
 
-# Example usage
-response = pipe("Summarize this: Climate change is...", max_new_tokens=150)
-print(response[0]['generated_text'])
-```
+response = pipe(
+    "Summarize this text: Artificial intelligence is...",
+    max_new_tokens=120
+)
 
-Option 2: Direct Downloads
+print(response[0]["generated_text"])
 
-· Android APK: GitHub Releases
-· Linux AppImage: Self-contained executable
-· Windows EXE: Coming soon
 
-📁 Model Details
 
-Training Information
 
-· Framework: Unsloth (4x faster training)
-· Method: Supervised Fine-Tuning (SFT)
-· Dataset: Custom multilingual mix with safety alignment
-· Hardware: Single RTX 4090 (24GB VRAM)
-· Training Time: ~8 hours
+🧪 Training Summary
 
-Safety & Limitations
 
-· ✅ Aligned with constitutional AI principles
-· ✅ Refuses harmful or unethical requests
-· ⚠️ Limited to 1B parameters – best for focused tasks
-· ⚠️ Not suitable for complex reasoning or creative writing
+Method: Supervised Fine-Tuning (SFT)
 
-🌍 Use Cases
+Data: Custom multilingual datasets with safety-focused filtering
 
-1. Developers: Local coding assistant without cloud costs
-2. Students: Study aid and document summarization
-3. Privacy-conscious users: AI that respects data boundaries
-4. Edge deployments: IoT and mobile applications
-5. Researchers: Baseline for lightweight model development
+Hardware: Single consumer GPU
 
-🔗 Community & Support
+Goal: Improve instruction-following, multilingual responses, and short-prompt behavior
 
-· GitHub Issues: Bug reports and feature requests
-· Discord Community: Live discussions and support
-· Hugging Face Spaces: Interactive demos and examples
-· Twitter/X: @dreamhubIR for updates
 
-📈 Roadmap
 
-· iOS version (Q4 2024)
-· Voice interface integration
-· Plugin system for extended functionality
-· Expanded language support (Arabic, Spanish)
-· Hardware acceleration benchmarks
 
-🙏 Acknowledgments
 
-· Google for the Gemma 3 base model
-· Unsloth for the amazing training optimizations
-· Hugging Face for the model hosting infrastructure
-· The 26 early adopters who helped validate WALL•E
+🛡 Safety & Limitations
+
+✅ Trained with safety-aware data
+✅ Avoids harmful or unethical requests
+⚠️ Limited reasoning depth due to 1B parameter size
+⚠️ Not intended for complex multi-step reasoning or creative writing
+
+
+
+🌍 Ideal Use Cases
+
+
+Local coding assistant
+
+Study and document summarization
+
+Privacy-focused users
+
+Lightweight edge deployments
+
+Research and experimentation with small LLMs
+
+
+
+
+
+🤝 Community & Links
+
+
+GitHub: https://github.com/unknownmsv/WALL-E
+
+Hugging Face Model: https://huggingface.co/sinamsv0/WALL-E
+
+Hugging Face Space: https://huggingface.co/spaces/sinamsv0/WALL-E-DEMO
+
+
+
+
+
+🔮 Roadmap (Planned)
+
+
+UI tools for local use
+
+Optional voice interface
+
+Extended language support
+
+Performance benchmarking on edge devices
+
 
----
 
-⭐ Star us on GitHub to support the project!
-🔗 Share your use cases and help improve WALL•E for everyone.
 
-"Small model, big impact – AI that works for you, not the other way around."
 
----
 
-#WALLE #Gemma3 #LocalAI #PrivacyAI #OpenSource #MobileAI #EdgeComputing #LightweightAI #MultilingualAI
+Small model, focused design.
+WALL•E proves that useful AI doesn’t have to be huge.